基于光子学的 HPC 资源分解.pdf

1、Efficient Intra-Rack Resource Disaggregation in HPC Using Co-Packaged DWDM PhotonicsGeorge Michelogiannakis,Yehia Arafa,Brandon Cook,Liang Yuan Dai,Abdel-Hameed Badawy,Madeleine Glick,Keren Bergman,John ShalfContact:miheloglbl.govPresentation Title|BERKELEY LABWhy Resource Disaggregation2Expensive r

2、esources are consistently underutilizedNERSCs PerlmutterJ Li et al.,“Analyzing Resource Utilization in an HPC System:A Case Study of NERSCs PerlmutterPresentation Title|BERKELEY LABIntra-rack resource disaggregation achieves most of the benefit with a fraction of the overhead3RAMCPUGPUSSDRAMRAMCPUGP

3、USSDRAMRAMCPUGPUSSDRAMRAMCPUGPUSSDRAMPresentation Title|BERKELEY LABTargets of Memory Disaggregation HardwareSatisfy each chips maximum escape bandwidthAchieve comparable BER with todays HPC systemsLess than 10-18We use forward error correction(FEC)and take into account its latencyMinimize energy ov

4、erheadMinimize latency overheadWill be imposed to latency-sensitive communication such as to and from memory4Our goalsPresentation Title|BERKELEY LABWith Emerging Co-Packaged PhotonicsPrototypes of dense wavelength division multiplexed links that are co-packaged to achieve their bandwidth densityBan

5、dwidths range from 100 Gbps to 2 TbpsRely on silicon comb laser sources5To maximize bandwidth density and meet goalsComb laser source providesmultiple frequenciesRings modulate differentfrequenciesPresentation Title|BERKELEY LABEmbedded Photonic Connectivity6CPUs,GPUs or routerComb sourcePresentatio

6、n Title|BERKELEY LABONIC PCIe-Driven Photonics Transceiver7Presentation Title|BERKELEY LABBuilding up a Rack Using MCMs8Demonstrated in 2.5D and 3D.Can use UCIe or CXLPresentation Title|BERKELEY LABPhotonic Switch At the Center of a Blade9Presentation Title|BE